Variable Selection for Multiply-Imputed Data

• Activity Number: 333
• Type: Topic Contributed
• Date/Time: Tuesday, August 3, 2010 : 10:30 AM to 12:20 PM
• Sponsor: Section on Survey Research Methods
• Author(s): Qixuan Chen*+ and Sijian Wang
• Companies: Columbia University and University of Wisconsin-Madison
• Keywords: multiple imputation ; regularization ; stepwise selection

Abstract

Multiple imputation is frequently used to handle the missing data. However, there are currently no guidelines for the extension of commonly used variable selection methods for complete data to the setting of multiply imputed data. We proposed two variable selection methods for multiply-imputed data. The first method is a modification of the traditional stepwise variable selection method, implemented by obtaining combined p-values using Rubin's multiple imputation combining rule first and then selecting variables base on combined p-values in each step of selection. The second method is based on the regularized joint-likelihood estimation approach. The coefficients of the same variable across all imputed data are treated as a group, and the group lasso penalty is applied for the purpose of variable selection. We demonstrated our methods using simulation studies and a real example.